Detecting a change inside a wind turbine is useful, but understanding why it has occurred is what allows engineers to decide what action to take.
That distinction sits at the heart of Liam Sharkey's view of artificial intelligence (AI) in wind operations and maintenance (O&M). The SVP of AI & Innovation at ONYX believes combining AI with engineering physics can give operators information they can interpret, verify and use when making maintenance decisions.
Moving beyond the black box
Wind operators manage mixed original equipment manufacturer (OEM) fleets, ageing assets, remaining useful life (RUL), derating decisions and equipment availability, while also dealing with weather and supply chain pressures.
Statistical AI models can help identify patterns, anomalies and potential mechanical problems. Their limitation is diagnosis. They may recognise that turbine behaviour is changing without explaining the physical mechanism responsible.
For engineers deciding whether to derate or shut down a turbine, that distinction matters.
Physics-informed AI combines data analysis with knowledge of structural dynamics, vibration and physical loading. The aim is to provide evidence engineers can use to verify a diagnosis rather than relying on statistical probability alone.
Reducing the risk of serious failures
Condition monitoring systems (CMS) can identify developing component problems, but deciding when intervention is necessary still depends heavily on engineering experience.
ONYX Insight says replacing a single wind turbine can cost more than $5 million, while business interruption can reach $100,000 per day in lost revenue.
AI could help turn fault information into better informed recommendations about when action is required.
Physics-informed models may be particularly useful for unusual failure modes because fundamental mechanical behaviour remains relevant even when a turbine behaves differently from examples contained within historical training data.
Useful from the beginning
New turbine models create another challenge. Statistical models may require at least 12 months of operating data to establish a useful anomaly-detection baseline, with longer histories needed for reliable fault diagnosis.
Physics-informed models can work with smaller or irregular datasets because they assess behaviour against established mechanical principles and wider industry failure experience.
This could allow operators to analyse new turbine models before extensive operating histories have accumulated.
Supporting decisions throughout asset life
As fleets age, operators increasingly need to understand how maintenance and operating decisions affect remaining useful life.
A turbine might, for example, be derated to extend operation until a planned service visit. Better forecasting can help determine whether such an approach is appropriate and reduce uncertainty around expensive interventions such as crane mobilisation.
Physics-informed AI can combine performance information, failure datasets and physical behaviour to support decisions from warranty periods through life extension, repowering and end-of-life assessment.
ONYX says its Cortex platform combines physics-based models derived from more than 50,000 failure events with AI.
Sharkey believes operators need to understand the reasoning behind AI recommendations, particularly when they could influence important engineering decisions.
When the decision could affect turbine availability, component life or a major repair, ‘the AI says so’ is not enough.
Read more in the complete PES Wind feature: https://pes.eu.com/exclusive-articles/why-wind-om-needs-ai-grounded-in-physics